public void Initialize(Context context, AcceleratorId acceleratorId, double[,] independents, double[] dependants)
{
AcceleratorId = acceleratorId;
AcceleratorType acceleratorType = AcceleratorId.AcceleratorType;
if (acceleratorType == AcceleratorType.CPU)
{
Accelerator = Accelerator.Create(context, AcceleratorId);
}
else if (acceleratorType == AcceleratorType.OpenCL)
{
Accelerator = CLAccelerator.Create(context, AcceleratorId);
}
else if (acceleratorType == AcceleratorType.Cuda)
{
Accelerator = CudaAccelerator.Create(context, AcceleratorId);
}
EvaluationKernel = Accelerator.LoadAutoGroupedStreamKernel <Index2, ArrayView2D <double>, ArrayView <double>, ArrayView <NodeGPU>, ArrayView <int>, ArrayView2D <double> >(EvaluationKernelFunction);
ProcessResultsKernel = Accelerator.LoadAutoGroupedStreamKernel <Index1, ArrayView2D <double>, ArrayView <double> >(ProcessResultsKernelFunction);
IndependentsTableSize = new Index2(independents.GetUpperBound(0) + 1, independents.GetUpperBound(1) + 1);
Independents = Accelerator.Allocate <double>(IndependentsTableSize);
Independents.CopyFrom(independents, new Index2(), new Index2(), IndependentsTableSize);
Dependants = Accelerator.Allocate <double>(dependants.Length);
Dependants.CopyFrom(dependants, 0, 0, dependants.Length);
}
public void GenerateHeightField(TerrainPatch patch)
{
using (var context = new Context())
{
AcceleratorId aid = Accelerator.Accelerators.Where(id => id.AcceleratorType == AcceleratorType.Cuda).FirstOrDefault();
if (aid.AcceleratorType != AcceleratorType.Cuda)
{
Console.WriteLine(@"There is no CUDA accelerator present. Doing nothing.");
return;
}
using (var accelerator = Accelerator.Create(context, aid))
using (var gpu_dem = accelerator.Allocate <short>(ViperEnvironment.Terrain.Data.Length))
using (var gpu_range = accelerator.Allocate <float>(cpu_range.Length))
using (var gpu_slope = accelerator.Allocate <float>(cpu_slope.Length))
using (var gpu_rise = accelerator.Allocate <short>(TerrainPatch.DefaultSize * TerrainPatch.DefaultSize))
{
gpu_dem.CopyFrom(ViperEnvironment.Terrain.Data, 0, 0, ViperEnvironment.Terrain.Data.Length);
gpu_range.CopyFrom(cpu_range, 0, 0, cpu_range.Length);
var launchDimension = new Index2(TerrainPatch.DefaultSize, TerrainPatch.DefaultSize);
var kernel1 = accelerator.LoadStreamKernel <Index2, ArrayView <short>, ArrayView <float>, ArrayView <short>, int, int>(RiseKernel1);
}
}
}
bool GetAccelerator(AcceleratorType t, out AcceleratorId aid)
{
aid = Accelerator.Accelerators.Where(id => id.AcceleratorType == t).FirstOrDefault();
if (aid.AcceleratorType != t)
{
Console.WriteLine(@"There is accelerator present of the desired type. Doing nothing.");
}
return(aid.AcceleratorType == t);
}
public Image<Rgba32> AddFilter(IImageFilter filter)
{
using (var context = new Context())
{
AcceleratorId acceleratorId = new AcceleratorId();
if (Accelerator.Accelerators.Where(a => a.AcceleratorType == AcceleratorType.Cuda).Count() != 0)
{
acceleratorId = Accelerator.Accelerators.Where(a =>
a.AcceleratorType == AcceleratorType.Cuda)
.FirstOrDefault();
}
else
{
acceleratorId = Accelerator.Accelerators.Where(a => a.AcceleratorType == AcceleratorType.CPU).FirstOrDefault();
}
using (var accelerator = Accelerator.Create(context, acceleratorId))
{
return filter.PerformFilter(accelerator);
}
}
}
static float EncodingToSlope(int encoding) => 4f * encoding / 1000000f; // Haven't looked at the numerics closely to see whether these are exact inverses. Don't care enough yet.
/// <summary>
/// Generate the near horizon for the patch, overwriting what was there.
/// The horizons will be in slope format, not angles
/// </summary>
/// <param name="target"></param>
public void AddNearHorizon(TerrainPatch target)
{
Debug.Assert(Terrain != null);
Debug.Assert(target != null);
if (target.Horizons == null)
{
target.InitializeHorizons();
}
using (var context = new Context())
{
AcceleratorId aid = Accelerator.Accelerators.Where(id => id.AcceleratorType == AcceleratorType.Cuda).FirstOrDefault();
//AcceleratorId aid = Accelerator.Accelerators.Where(id => id.AcceleratorType == AcceleratorType.CPU).FirstOrDefault();
if (aid.AcceleratorType != AcceleratorType.Cuda)
{
Console.WriteLine(@"There is no CUDA accelerator present. Doing nothing.");
return;
}
using (var accelerator = Accelerator.Create(context, aid))
{
target.Matrices = null; // Be sure!!
target.FillPoints(Terrain);
target.FillMatricesRelativeToPoint(Terrain, target.Points[0][0]);
// Matrices
var cpu_matrices_size = target.Height * target.Width * 12;
var basePoint = target.Points[0][0];
var cpu_matrices = MakeCPUMatrices(target);
// Horizon (load from target)
var cpu_horizon_size = target.Height * target.Width * Horizon.HorizonSamples;
var cpu_horizon = new float[cpu_horizon_size];
// Initialize to float.MinValue. Maintain horizon as slope
for (var i = 0; i < cpu_horizon_size; i++)
{
cpu_horizon[i] = float.MinValue;
}
// Caster array
const int dem_size = TerrainPatch.DEM_size;
const int patch_size = TerrainPatch.DefaultSize;
var maxDistance = (float)TerrainPatch.MaximumLocalDistance;
var border = 2 + (int)Math.Ceiling(maxDistance); // the 2 is margin for the bilinear interpolation
var line_min = Math.Max(0, target.Line - border);
var line_max = Math.Min(dem_size - 1, target.Line + patch_size + border); // 1 more than the highest index
var line_size = line_max - line_min;
var line_offset = target.Line - line_min;
var sample_min = Math.Max(0, target.Sample - border);
var sample_max = Math.Min(dem_size - 1, target.Sample + patch_size + border); // 1 more than the highest index
var sample_size = sample_max - sample_min;
var sample_offset = target.Sample - sample_min;
var cpu_caster_points_size = line_size * sample_size * 3;
var cpu_caster_points = new float[cpu_caster_points_size];
FillNearCasterArray(cpu_caster_points, target.Points[0][0], line_min, line_max, sample_min, sample_max);
//DumpNearFieldTestCsv(cpu_caster_points, cpu_matrices);
var test_floats_size = Horizon.HorizonSamples * TerrainPatch.NearHorizonOversample;
var test_floats = new float[test_floats_size];
using (var gpu_matrices = accelerator.Allocate <float>(cpu_matrices_size))
using (var gpu_horizon = accelerator.Allocate <float>(cpu_horizon_size))
using (var gpu_caster_points = accelerator.Allocate <float>(cpu_caster_points_size))
using (var gpu_test_floats = accelerator.Allocate <float>(test_floats_size))
{
gpu_matrices.CopyFrom(cpu_matrices, 0, 0, cpu_matrices_size);
gpu_horizon.CopyFrom(cpu_horizon, 0, 0, cpu_horizon_size);
gpu_caster_points.CopyFrom(cpu_caster_points, 0, 0, cpu_caster_points_size);
gpu_test_floats.CopyFrom(test_floats, 0, 0, test_floats_size);
var groupSize = accelerator.MaxNumThreadsPerGroup;
Index launchDimension = Horizon.HorizonSamples * TerrainPatch.NearHorizonOversample;
const float d_min = 1f;
var d_max = (float)TerrainPatch.MaximumLocalDistance;
const float d_step = TerrainPatch.LocalStep;
var kernel1 = accelerator.LoadAutoGroupedStreamKernel <Index, int, int, int, int, int, int, float, float, float, ArrayView <float>, ArrayView <float>, ArrayView <float>, ArrayView <float> >(NearHorizonKernel1);
Console.WriteLine(@"Launching near horizon kernels ... ");
if (true)
{
for (var target_line = 0; target_line < TerrainPatch.DefaultSize; target_line++)
{
for (var target_sample = 0; target_sample < TerrainPatch.DefaultSize; target_sample++)
{
kernel1(launchDimension, target_line, target_sample, line_offset, sample_offset, line_size, sample_size, d_min, d_max, d_step, gpu_caster_points, gpu_matrices, gpu_horizon, gpu_test_floats);
}
}
}
else
{
kernel1(launchDimension, 0, 0, line_offset, sample_offset, line_size, sample_size, d_min, d_max, d_step, gpu_caster_points, gpu_matrices, gpu_horizon, gpu_test_floats);
}
gpu_horizon.CopyTo(cpu_horizon, 0, 0, cpu_horizon_size);
gpu_test_floats.CopyTo(test_floats, 0, 0, test_floats_size);
// Update the horizons
for (var line = 0; line < TerrainPatch.DefaultSize; line++)
{
for (var sample = 0; sample < TerrainPatch.DefaultSize; sample++)
{
var offset = (line * TerrainPatch.DefaultSize + sample) * Horizon.HorizonSamples;
var buffer = target.Horizons[line][sample].Buffer;
for (var i = 0; i < Horizon.HorizonSamples; i++)
{
buffer[i] = cpu_horizon[i + offset];
}
}
}
}
}
}
}
/// <summary>
/// Update the horizons of a patch based on a list of shadow casters.
/// The horizons will be in slope, not angle, format
/// </summary>
/// <param name="target"></param>
/// <param name="casters"></param>
public void UpdateHorizons(TerrainPatch target, List <TerrainPatch> casters)
{
Debug.Assert(Terrain != null);
if (casters.Count < 1)
{
return;
}
using (var context = new Context())
{
AcceleratorId aid = Accelerator.Accelerators.Where(id => id.AcceleratorType == AcceleratorType.Cuda).FirstOrDefault();
if (aid.AcceleratorType != AcceleratorType.Cuda)
{
Console.WriteLine(@"There is no CUDA accelerator present. Doing nothing.");
return;
}
using (var accelerator = Accelerator.Create(context, aid))
{
target.FillPoints(Terrain);
target.FillMatricesRelativeToPoint(Terrain, target.Points[0][0]);
// Matrices
var cpu_matrices_size = target.Height * target.Width * 12;
var basePoint = target.Points[0][0];
var cpu_matrices = MakeCPUMatrices(target);
// Horizon (load from target)
var cpu_horizon_size = target.Height * target.Width * Horizon.HorizonSamples;
var cpu_horizon = new int[cpu_horizon_size];
for (var line = 0; line < TerrainPatch.DefaultSize; line++)
{
for (var sample = 0; sample < TerrainPatch.DefaultSize; sample++)
{
var offset = (line * TerrainPatch.DefaultSize + sample) * Horizon.HorizonSamples;
var buffer = target.Horizons[line][sample].Buffer;
for (var i = 0; i < Horizon.HorizonSamples; i++)
{
cpu_horizon[i + offset] = SlopeToEncoding(buffer[i]);
}
}
}
// Caster points
var cpu_caster_points_size = casters[0].Width * casters[0].Height * 3;
var cpu_caster_points = new float[cpu_caster_points_size];
// test array
var cpu_test_array = new float[20];
using (var gpu_matrices = accelerator.Allocate <float>(cpu_matrices_size))
using (var gpu_horizon = accelerator.Allocate <int>(cpu_horizon_size))
using (var gpu_caster_points = accelerator.Allocate <float>(cpu_caster_points_size))
using (var gpu_test_array = accelerator.Allocate <float>(cpu_test_array.Length))
{
gpu_matrices.CopyFrom(cpu_matrices, 0, 0, cpu_matrices_size);
gpu_horizon.CopyFrom(cpu_horizon, 0, 0, cpu_horizon_size);
var groupSize = accelerator.MaxNumThreadsPerGroup;
var launchDimension = new GroupedIndex2(
new Index2(128, 128), // (data.Length + groupSize - 1) / groupSize, // Compute the number of groups (round up)
new Index2(1, 128));
var kernel1 = accelerator.LoadSharedMemoryStreamKernel1 <GroupedIndex2, ArrayView <float>, ArrayView <float>, ArrayView <int>, ArrayView <float>, ArrayView <int> >(ShadowKernel1);
//var stopwatch = new Stopwatch();
//stopwatch.Start();
foreach (var caster in casters)
{
caster.FillPoints(Terrain);
CopyPointsToCpuArray(caster, basePoint, cpu_caster_points);
gpu_caster_points.CopyFrom(cpu_caster_points, 0, 0, cpu_caster_points_size);
kernel1(launchDimension, gpu_caster_points, gpu_matrices, gpu_horizon, gpu_test_array);
accelerator.Synchronize();
}
// Copy out data
gpu_horizon.CopyTo(cpu_horizon, 0, 0, cpu_horizon_size);
gpu_test_array.CopyTo(cpu_test_array, 0, 0, cpu_test_array.Length);
//stopwatch.Stop();
//Console.WriteLine($"kernel time={stopwatch.Elapsed} cpu_horizon.Max()={cpu_horizon.Max()} cpu_horizon[0]={cpu_horizon[0]}");
// Update the horizons
for (var line = 0; line < TerrainPatch.DefaultSize; line++)
{
for (var sample = 0; sample < TerrainPatch.DefaultSize; sample++)
{
var offset = (line * TerrainPatch.DefaultSize + sample) * Horizon.HorizonSamples;
var buffer = target.Horizons[line][sample].Buffer;
for (var i = 0; i < Horizon.HorizonSamples; i++)
{
buffer[i] = EncodingToSlope(cpu_horizon[i + offset]);
}
}
}
//Console.WriteLine($" max slope={cpu_horizon.Select(EncodingToSlope).Max()}");
}
}
}
}